300 Thomas C. Purnell
Since some of the third group’s data were presented for an identi¿ cation
evaluation in a forced-choice perceptual experiment (Purnell et al. 2005a),
we compare the acoustic and perceptual data where appropriate. Nineteen
subjects listened to the data, responding to the ¿ nal sound in each word they
heard. Only three of the listeners did not report a dialect region in Wisconsin
or southern Minnesota. The error rates for the group 3 tokens which were
considered important were those over 10%: /t/ as [d], 10%; /s/ as [t], 11%; /s/
as [z], 18%; and /z/ as [s], 33%. Thus, a parallel analysis is conducted on the
group 3 data that was presented in the perception experiment. The voicing
category is transformed based on the error rates over 10%, i.e., if a /z/ target
token was perceived as either [t] or [s] more than 10% of the time, then it was
coded as voiceless. There are three possible expectations with respect to this
data. First, we might expect that the subjects will use a Wisconsin English set
of weighted measures, most like group 4 given the proximity in age. Second,
we could expect that the listeners use a Wisconsin English set of weighted
measures parallel to group 3, given that this is the input to the experiment.
Third, we might expect that the listeners use a default, network English set of
weighted measures, i.e., more like the controls in group 5.
3.3 Results
Results of the stepwise and canonical discriminant analyses are shown in
Tables 13.3 through 13.7. The stepwise analysis results are shown in Table
13.3. Per the stepwise process outlined previously, for each group the four
strongest measures are listed in descending order of the degree to which they
contribute to the overall strength of the statistical model for that group. For
example, percent glottal pulsing, vowel duration, change in F1, and change in
F0 are the four measures for group 1 that show relative statistical signi¿ cance.
They are listed in their present order, not because of their univariate signi¿ -
cance (compare vowel duration, p=0.0787, with change in F1, p=0.0756), but
because of their contribution to the squared canonical correlation value (a
resultant 0.485 and 0.523, respectively). Each line of data on Table 13.3 rep-
resents the ¿ nal model output given the addition of the variable on that line
along with any variable already added to the model. Of interest is the value
for the average squared canonical correlation (ASCC) on the last line for each
group because this value depicts the degree of well-separated discriminant
space and approximates the amount of variation accounted for by the mea-
sures accepted into the model. For the ASCC values, a value of 1 represents
the greatest distinction between groups, and thus how separate voiced tokens